What the AI moderator can't do

The AI moderator handles most interviews end-to-end. Here are the places it'll struggle, and what to do about them.

When this matters

If your study depends on something on this list, design around it before publishing, not after the first 10 interviews come back empty.

Things it can't (or won't) do

• Customize the greeting or the goodbye. Both are built in and can't be edited or removed. To add framing in your own words, put a Statement as the first item (it plays right after the greeting) or the last item (just before the goodbye).
• Skip questions on its own. Even if a participant covers a later question in an earlier answer, the moderator will still ask the later one unless you've put it inside a conditional section. Use conditional sections to skip pre-empted questions.
• Branch the guide on what a participant said in conversation. Conditional sections work on structured answers only (screener answers, earlier survey answers, or quotas). For free speech, capture the driver as a survey question first, then gate later sections on that answer. Within a question, follow-ups do react to what was said. That's what Follow-up guidance is for.
• Auto-advance on a timer. Items advance when the conversation completes or the participant acts. For activities with a natural finish line, use a goal-based Figma prototype; for pacing control, hand advancement to the participant via Question flow → Participant-controlled.
• Script the exact wording of follow-ups. Follow-up wording is composed live. You control how many (Follow-up depth) and what to probe (Follow-up guidance). If a question must be asked in specific words, make it its own guide item with Ask question verbatim.
• Skip past a survey question or answer it for the participant. Survey questions wait for the participant. If you want it skippable, make it an open question instead, or gate it with a conditional section so some participants never see it.
• Switch language mid-call. The interview runs in the language the participant chose at the start. The moderator can recognize individual words in another language (brand names, proper nouns) but won't change the conversation language partway through.
• Hand-pick which participant gets which variant. Variant assignment is automatic and random. If specific participants need specific content, use conditional sections keyed to screener answers or quotas instead of randomization.
• Interrupt mid-thought reliably. It waits for natural pauses. Participants who talk in long uninterrupted paragraphs may finish a thought before the moderator can redirect.
• Make judgments your guide hasn't asked for. It won't decide a participant is "lying" or "biased" and probe on its own. It follows your guide.
• See what's on screen by default. Agent vision (beta) lets the moderator look at the camera and shared screen periodically. Without it, the moderator works from what the participant says, which is why it asks them to confirm they can see media before moving on.
• Read documents the participant uploads mid-call. Ask them to share their screen with the document open instead.
• Run past your Session timeout for an unresponsive participant. If nothing happens for the configured timeout (30 sec to 15 min), the call ends after a 30-second warning. Anything the moderator says or asks resets the clock. On a Task, where the moderator stays quiet while the participant explores on their own, the participant talking resets it too, so thinking out loud won't end the call. On other elements talking alone doesn't reset it, so an interview that has genuinely stalled still closes out.

Related